Chebyshev approximation represents a function on a finite interval using specially structured polynomials. Its practical advantage is that a smooth function can often be represented accurately with relatively few coefficients, while cosine-spaced sampling and recurrence-based evaluation help control numerical difficulties. The gold standard is an approximation whose interval, coefficient convention and error test are explicit, and whose accuracy is checked at inputs that did not merely construct the curve.
This guide explains how that standard works through calculations you can reproduce. You will build Chebyshev polynomials from a cosine identity, transform a physical interval, reconstruct a function from samples, evaluate a series without expanding powers, and diagnose aliasing and misleading error claims. The worked functions are deliberately constructed examples, not experimental findings. For the broader learning purpose, begin with the Gold Standard master and the Education parent.
- Start with a stated interval and a clear absolute or relative error requirement.
- Keep the polynomial basis, sample grid and approximation objective separate when comparing methods.
- Check interval scaling and coefficient normalization with examples whose answers are known independently.
- Inspect off-grid errors and unresolved oscillations before trusting a small coefficient tail.
- Validate derivatives, integrals and extrapolation separately when those operations matter to the final application.
Choose your Chebyshev approximation route
Chapters 1–3: Understand the basis and interval
Chapters 4–6: Build and evaluate approximations
Chapters 7–9: Measure error and detect hidden failures
Chapters 10–12: Validate, repair and apply
Chapter index
Understand the basis and interval · Chapters 1–3
1. The problem: a useful function at lower cost
Build and evaluate approximations · Chapters 4–6
4. Choosing cosine-spaced samples
Measure error and detect hidden failures · Chapters 7–9
7. Three meanings of approximation error
Validate, repair and apply · Chapters 10–12
10. Smoothness, corners and extrapolation
Transfer and continue learning · Chapters 13–14
CHAPTER 1 OF 14 · Understand the basis and interval
1. The problem: a useful function at lower cost
Suppose evaluating a function requires a long simulation, a numerical integral or a complicated formula. If you need its value thousands of times over the same interval, a short polynomial representation may be cheaper. The approximation replaces repeated expensive work with a small amount of arithmetic, provided the resulting error is acceptable for the task.
The problem has several parts. You need a domain, a function or dataset, a permitted polynomial degree, an error measure and a way to evaluate the answer. Choosing Chebyshev polynomials addresses the representation. Choosing Chebyshev nodes addresses sampling. Neither choice alone specifies whether you are interpolating values, projecting a function into a polynomial space, fitting noisy measurements or minimizing the largest error.
Separate the objects before comparing methods
A basis is a collection of building blocks. The usual power basis is 1, t, t² and so on. A Chebyshev basis spans the same degree-limited polynomial space, but its building blocks have different shapes. Changing the basis does not enlarge that space. It changes the coordinates in which you describe and compute the polynomial.
Nodes are the inputs at which values are sampled. You can represent the interpolating polynomial in a Chebyshev basis even when its nodes are not Chebyshev nodes. Conversely, you can interpolate at Chebyshev nodes and then expand the answer into ordinary powers. The mathematical polynomial may remain the same while its numerical handling changes.
Finally, an objective tells you what “best” means. Passing through samples, minimizing average squared error and minimizing maximum error are different objectives. They can produce different polynomials of the same degree from the same target function.
Before calculating, write a sentence such as: “Approximate this known function on this closed interval with a degree-four polynomial and keep absolute error below this tolerance.” That sentence exposes missing information immediately. If the input is noisy measurements rather than an exactly evaluable function, record that distinction too. An approximation cannot recover detail that the observations never resolved.
CHAPTER 2 OF 14 · Understand the basis and interval
2. From cosine waves to polynomial building blocks
The first-kind Chebyshev polynomial Tₖ is defined on [−1,1] by Tₖ(cos θ) = cos(kθ). Although this definition uses trigonometry, each Tₖ is a polynomial in its argument. The first two are T₀(t) = 1 and T₁(t) = t. The cosine addition identity gives Tₖ₊₁(t) = 2tTₖ(t) − Tₖ₋₁(t).
Apply the recurrence once to obtain T₂(t) = 2t²−1. Apply it again to obtain T₃(t) = 4t³−3t. The next polynomial is T₄(t) = 8t⁴−8t²+1. These expressions can be checked at t = 1: every one equals one, as the cosine definition predicts.
Why the interval matters
For t in [−1,1], write t = cos θ with real θ. Then |Tₖ(t)| ≤ 1 because a cosine has magnitude at most one. This boundedness gives a direct way to control the effect of coefficient errors. Outside that interval the same polynomial is still defined, but the bound no longer applies.
At t = 2, for example, T₀ = 1, T₁ = 2, T₂ = 7, T₃ = 26 and T₄ = 97. A coefficient that contributes at most one hundredth inside the standard interval could contribute almost one at t = 2 when it multiplies T₄. Extrapolation is therefore a different accuracy problem.
The recurrence also reveals parity. T₀ is even, T₁ is odd, and subsequent polynomials alternate between even and odd. An even target function can be represented using even-indexed terms when the representation respects that symmetry. An odd target similarly uses odd-indexed terms.
The NIST reference on approximation techniques supplies the defining identity, recurrence and orthogonality relations. The calculations here use the convention p(t) = c₀ + c₁T₁(t) + …, with the constant coefficient included in full. Some references use a halved constant term; copying coefficients between those conventions without conversion changes the function.
CHAPTER 3 OF 14 · Understand the basis and interval
3. Mapping the physical interval correctly
Chebyshev formulas are most naturally expressed on [−1,1]. To approximate a function of x on [a,b], where a < b, define t = (2x−a−b)/(b−a). The inverse is x = (a+b)/2 + (b−a)t/2. These formulas send the left endpoint to minus one, the midpoint to zero and the right endpoint to one.
Suppose x is a temperature in degrees Celsius over [20,80]. Then t = (x−50)/30. At 20°C, t = −1; at 50°C, t = 0; at 80°C, t = 1. A model stored as p(t) must receive this transformed coordinate, not the raw temperature.
A complete numerical example
Take p(t) = 4 + 2T₁(t) + T₂(t). In ordinary powers this is 3+2t+2t². At x = 65°C, t = 1/2, so p = 3+1+1/2 = 4.5. Evaluating the same coefficient array at t = 65 would not be a minor rounding mistake; it would evaluate an entirely different input far outside the specified domain.
Changing input units also changes the mapping. If the physical temperatures are instead recorded in kelvin, both interval endpoints and every query must be converted consistently. The dimensionless t for the same physical state remains unchanged. This is a useful test because it checks the interpretation of the model rather than only the arithmetic of its evaluator.
Store the endpoints with the coefficients. A coefficient array alone does not identify the physical function. Store the output units as well, and distinguish an absolute error tolerance in those units from a dimensionless relative tolerance.
For a software implementation, test the three landmarks before testing random inputs. Endpoint and midpoint errors often identify a reversed interval, incorrect offset or missing factor of two. If the model is later moved to another interval, decide whether it represents a newly transformed function or whether its coefficients must be recomputed. Simply changing the stored endpoints generally changes what the model means.
CHAPTER 4 OF 14 · Build and evaluate approximations
4. Choosing cosine-spaced samples
For N sample points of the first kind, use tⱼ = cos((j+1/2)π/N), with j from zero to N−1. These are the roots of the degree-N Chebyshev polynomial. They lie strictly inside [−1,1]. The corresponding angles are evenly spaced, but their projections onto the t-axis cluster near the endpoints.
For N = 3, the points are √3/2, zero and −√3/2. Numerically the outer points are approximately ±0.866025. Their order in the formula runs from right to left; sorting them into ascending order is fine if the sampled values are reordered with them.
Roots and extrema are different grids
Another common choice uses tⱼ = cos(jπ/n), with j from zero to n. This grid has n+1 points and includes both endpoints. For n = 4 the points are one, √2/2, zero, −√2/2 and minus one. They are extrema of T₄. Formula conventions, transform normalization and endpoint weights differ between these grids.
Why cluster near the ends? Global polynomial interpolation can be sensitive to node placement, especially near interval boundaries. Cosine-spaced nodes are designed to avoid the particularly poor behaviour that can occur with high-degree interpolation at equally spaced nodes. They are not a guarantee that every continuous function will be accurately resolved at a chosen finite degree.
If the data already exist at fixed measurement locations, you cannot pretend that they were collected on a different grid. Resampling those measurements first introduces another approximation. In that setting, a least-squares fit in a Chebyshev basis may be appropriate, but its uncertainties and objective differ from direct interpolation of an evaluable function at chosen nodes.
The NumPy documentation for chebinterpolate identifies its sampling convention as first-kind points on [−1,1]. Read that convention before comparing its coefficients with another routine. Two different grids may legitimately produce different finite-degree approximations even when both routines work correctly.
CHAPTER 5 OF 14 · Build and evaluate approximations
5. Recovering coefficients from a small dataset
Use the first-kind points with N = 3 and the constructed function g(t) = 1+2t+t². Its exact Chebyshev representation follows immediately from t² = (T₂+T₀)/2: g(t) = 3/2 + 2T₁(t) + T₂(t)/2. We will recover those coefficients from the three samples rather than from symbolic expansion.
Let s = √3/2. The values are g(s) = 7/4+√3, g(0) = 1 and g(−s) = 7/4−√3. The mean of these three values is 3/2, giving the constant coefficient. For this first-kind grid, the other coefficients are obtained by multiplying sampled values by cos(kθⱼ), summing and multiplying by 2/N.
For k = 1, the cosine values are s, zero and −s. The weighted sum is s[g(s)−g(−s)] = s(2√3) = 3. Multiplying by 2/3 gives c₁ = 2. For k = 2, the cosine values are 1/2, −1 and 1/2. The weighted sum is (g(s)+g(−s))/2−1 = 3/4. Multiplying by 2/3 gives c₂ = 1/2.
What this reconstruction proves
The recovered polynomial exactly reproduces this degree-two target in exact arithmetic. That is a controlled test of the grid, normalization and coefficient ordering. It does not show that three samples will capture an arbitrary function. Higher-order behaviour can be invisible or misidentified on a small grid.
At t = 1/2, evaluate the recovered series: 3/2 + 2(1/2) + (1/2)(−1/2) = 9/4. Direct substitution into 1+2t+t² gives the same result. This off-grid test checks more than reproducing the original three values.
In a larger implementation, a cosine transform computes these sums efficiently. Its scaling convention must still match the series definition. Test a constant, a single basis polynomial and a mixed low-degree polynomial before trusting the transform on a function whose answer you cannot inspect by hand.
CHAPTER 6 OF 14 · Build and evaluate approximations
6. Evaluating a series with Clenshaw recurrence
Once coefficients are available, repeatedly expanding every Tₖ into powers is unnecessary. For p(t) = c₀ + Σ cₖTₖ(t), initialize bₙ₊₁ = bₙ₊₂ = 0. Work backward for k = n down to one using bₖ = cₖ+2tbₖ₊₁−bₖ₊₂. Finish with p(t) = c₀+tb₁−b₂.
Take p(t) = 2−T₁(t)+T₂(t)/2+T₃(t)/4, and evaluate at t = 1/2. Start with b₄ = b₅ = 0. Then b₃ = 1/4, b₂ = 1/2+1/4 = 3/4, and b₁ = −1+3/4−1/4 = −1/2. The final result is 2+(1/2)(−1/2)−3/4 = 1.
Check independently: at one half, T₁ = 1/2, T₂ = −1/2 and T₃ = −1. Direct substitution gives 2−1/2−1/4−1/4 = 1. Agreement checks both the recurrence and its final step.
Efficient evaluation still needs a contract
The recurrence takes work proportional to the number of coefficients and keeps only a small amount of working state. It avoids explicitly forming every expanded polynomial. However, it does not remove sensitivity inherent in a nearly cancelling sum, an inaccurate argument or extrapolation far outside the interval.
The existing Clenshaw recurrence specialist article develops this evaluator in more detail. The official NumPy chebval reference specifies coefficients in increasing Chebyshev degree and uses the full constant term convention adopted here.
Keep the representation in its natural basis unless another operation requires conversion. A correct conversion to power coefficients can still create a less convenient numerical representation at high degree. If two evaluators disagree, compare on a low-degree exact example and then increase working precision. Separate coefficient-generation error, conversion error and evaluation error instead of attributing every discrepancy to the approximation itself.
CHAPTER 7 OF 14 · Measure error and detect hidden failures
7. Three meanings of approximation error
Interpolation error measures the difference between the target and a polynomial constrained to match chosen sample values. A truncated series instead keeps only a selected initial set of expansion coefficients. A minimax polynomial minimizes the largest absolute error over the whole interval among polynomials of the permitted degree. These constructions can be close, but they are not identical definitions.
Consider f(t) = t⁴ on [−1,1]. From the expression for T₄ and the identity for t², derive t⁴ = 3/8 + T₂(t)/2 + T₄(t)/8. Truncating after degree two gives p(t) = 3/8+T₂(t)/2 = t²−1/8. The error is T₄(t)/8, so its maximum absolute value is exactly 1/8.
At t = −1, −1/√2, zero, 1/√2 and one, that error alternates between +1/8 and −1/8. This example also has the minimax alternation structure. It is a favourable special case; it does not establish that arbitrary truncated Chebyshev expansions are always minimax.
A direct comparison with interpolation
Interpolate t⁴ at the three first-kind nodes from chapter four. At the two nonzero nodes, t² = 3/4 and t⁴ = 9/16; at zero both vanish. The interpolating quadratic is q(t) = 3t²/4. At either endpoint its error is 1/4, twice the maximum error of p.
To confirm that 1/4 is q’s maximum absolute error, put u = t² in [0,1]. The error becomes u²−3u/4. Its interior minimum occurs at u = 3/8 and equals −9/64. Its endpoint values are zero and 1/4. Therefore the largest magnitude is 1/4.
The Remez exchange and minimax specialist article explains how an algorithm targets the uniform-error objective. Use it when that objective matters, rather than relabelling every Chebyshev interpolant as an exact minimax solution.
CHAPTER 8 OF 14 · Measure error and detect hidden failures
8. Turning coefficient information into bounds
Because |Tₖ(t)| ≤ 1 on the standard interval, a finite omitted tail obeys |Σ cₖTₖ(t)| ≤ Σ|cₖ|. This follows directly from the triangle inequality. For an infinite series, the same argument applies when the tail is absolutely summable and represents the target as assumed.
Suppose a constructed model has only three omitted terms: 0.008T₄−0.003T₅+0.001T₆. Their total contribution has magnitude at most 0.012 throughout [−1,1]. This bound may be conservative because the terms need not reach their largest magnitudes with matching signs at the same point. It is nevertheless a valid bound under the stated finite-tail assumption.
If the retained coefficients each have known absolute errors δₖ, their combined contribution to function-value uncertainty is bounded by Σ|δₖ| on the same interval. Add separate evaluation and argument-error allowances when they matter. Coefficient uncertainty is only one part of the budget.
An estimate is not automatically a certificate
In practice, the coefficients beyond the computed degree are usually unknown. Observing that the last few computed coefficients are small suggests that the function may be well resolved, but it does not bound an unseen tail without additional assumptions. Aliasing can even make computed high-order coefficients vanish while a missed component remains large between nodes.
A rigorous tail estimate requires information such as a proved coefficient bound or an appropriate regularity argument. A practical estimate can combine coefficient decay, degree refinement and independent pointwise tests, provided it is reported as an estimate rather than a universal guarantee.
Choose absolute and relative errors deliberately. Relative error is problematic near a target zero: a tiny absolute discrepancy can correspond to a huge relative discrepancy or an undefined ratio. One useful acceptance rule compares |f−p| with an absolute allowance plus a relative allowance times |f|. The values of those allowances must come from the application, not from the number of digits a computer happens to print.
CHAPTER 9 OF 14 · Measure error and detect hidden failures
9. Aliasing: when samples conceal a component
Sampling cannot identify every possible function from finitely many values. Chebyshev grids have a particularly clear demonstration. At the N first-kind nodes tⱼ = cos((j+1/2)π/N), the polynomial Tₙ with index N vanishes, because cos((j+1/2)π) = 0.
For N = 3, suppose the target is f(t) = T₃(t). All three sampled values are zero. A degree-two interpolant built from those values is the zero polynomial. Yet f(1) = 1 and f(−1) = −1. The samples were exact; the reconstruction still missed the function between and beyond the sampled nodes inside the interval.
A false constant can also appear
At the same first-kind nodes, T₂ₙ(tⱼ) = cos((2j+1)π) = −1. Therefore a target T₆ sampled at three nodes looks exactly like the constant minus one. At t = 1, however, T₆(1) = 1. The resulting error is two at that endpoint.
These examples explain why checking only the construction nodes is insufficient, and why a small computed coefficient tail can mislead. The data can project unresolved variation into lower-degree terms. The error is a lack of sampling resolution, not necessarily a bug in the transform.
Use a larger grid, include independent off-grid queries and compare successive representations. No finite set of tests proves accuracy for an unrestricted unknown function, but tests informed by smoothness and application knowledge can provide much stronger evidence than construction residuals alone.
The spectral methods article connects this issue to global basis methods and unresolved modes. Keep the lesson specific: increasing polynomial degree helps only when the additional samples or other information actually resolve the missing behaviour. Reusing the same insufficient information in a larger coefficient array does not create new evidence.
CHAPTER 10 OF 14 · Validate, repair and apply
10. Smoothness, corners and extrapolation
Rapid approximation is closely related to regularity. A function that is analytic in a suitable neighbourhood of the interval can often be approximated with geometrically decreasing errors as degree increases. The location of nearby complex singularities affects that rate. Smoothness visible on a real plot alone does not provide a numerical convergence constant.
A corner changes the situation. The function |t| is continuous but has a derivative discontinuity at zero. Its global polynomial approximations generally require more terms for a given accuracy than comparably scaled analytic examples. A jump discontinuity is more severe: a continuous polynomial cannot converge uniformly to a discontinuous target on an interval containing the jump.
A short reason for the discontinuity limitation
Suppose continuous approximants converged uniformly to a function with a jump. Uniform convergence would preserve continuity of the limit, contradicting the jump. This is a structural obstacle, not a failure that can be removed merely by adding enough terms to one global polynomial.
When a known change of regime occurs, split the interval and approximate each appropriate piece. Record how the pieces meet and whether the application needs matching values or derivatives. If the regime boundary is uncertain, acknowledge that uncertainty rather than selecting a convenient breakpoint only because it improves a plot.
Extrapolation deserves its own test. The boundedness of Tₖ inside [−1,1] disappears outside it, as chapter two showed numerically. Small coefficient errors and small high-degree terms can become large. A good in-interval approximation may therefore be an unreliable predictor even a modest distance beyond its declared range.
The Rational Interpolation root offers another representation when a ratio of polynomials fits the structure better. That alternative introduces denominator and pole checks. The useful decision is not whether one family is universally superior, but which family matches the available information and remains verifiable over the intended domain.
Start with data and coordinate checks. Are the inputs finite, correctly ordered and expressed in the declared units? Are the interval endpoints distinct? Were paired values reordered together? Does evaluating the model at the physical midpoint correspond to evaluating its basis at zero? These checks often explain a large error faster than increasing the degree.
Next check conventions. A constant function should produce only the constant coefficient, apart from floating-point noise. A single Tₖ should be reproduced when the degree and sampling resolution suffice. The mixed polynomial from chapter five checks both parity and the factor used for nonconstant coefficients.
Diagnose by the pattern of failure
If node values are wrong, investigate construction or normalization. If nodes match but independent values fail, investigate resolution and the chosen degree. If evaluation changes dramatically when the representation is converted to powers, investigate conversion and arithmetic. If only physical-unit queries fail, investigate the interval map.
If increasing degree reduces error initially and then stops helping, compare the plateau with measurement uncertainty and machine precision. More terms may begin fitting noise or amplifying arithmetic difficulties. Use repeated observations or a lower-degree fit when that better matches what the data can support.
If a narrow region dominates the error, inspect local regularity and consider interval splitting. If the approximation violates a physical constraint such as positivity, do not assume the Chebyshev basis should have preserved it. Generic polynomial approximation does not enforce every property of the sampled values. Add appropriate constraints or choose a method designed for the required property.
After a repair, repeat the test that originally failed and the earlier tests that passed. Retain a small set of exact examples as regression checks. Record actual maximum sampled errors and where they occurred, while keeping sampled maxima distinct from mathematically certified bounds.
Finally, preserve enough information to reproduce the result: source function or dataset version, interval, units, grid convention, degree, coefficient normalization, evaluator and acceptance tests. A visually smooth curve without those details is difficult to assess and even harder to maintain when its operating range changes.
CHAPTER 12 OF 14 · Validate, repair and apply
12. Derivatives, integrals and practical use
A compact approximation can accelerate repeated function evaluation, but downstream operations need separate accuracy checks. Differentiation can magnify high-degree components that are small in value. For example, Tₙ′(1) = n². An omitted term 0.0001T₁₀ contributes at most 0.0001 to function values on [−1,1], yet contributes 0.01 to the derivative at the right endpoint.
On a physical interval [a,b], the derivative with respect to x includes the factor 2/(b−a). A narrow interval can therefore make a given error in the dimensionless derivative more important in physical units. Forgetting the chain-rule factor can invalidate an otherwise accurate polynomial representation.
Integrating the same representation
Integration also requires the coordinate change: dx = (b−a)dt/2. Consider p(t) = 4+2T₁(t)+T₂(t) = 3+2t+2t². Its integral over [−1,1] is 6+4/3 = 22/3. Over the temperature interval [20,80], the corresponding integral with respect to x is 30 times that value, or 220 in output-units times degrees Celsius.
If a uniform function-value error bound E is established over [a,b], the absolute integral error is at most (b−a)E. This follows by integrating the absolute error bound. The statement assumes an ordinary integral of the target and approximant; it does not certify unrelated quantities such as the location of a root.
For root finding, a small value error near a nearly flat crossing can cause a large location error. For optimization, the accuracy of a minimum and its location may depend on derivatives and curvature. Validate the quantity the eventual application uses, not merely a convenient proxy.
The How X Works hub connects these mathematical choices to wider mechanisms and systems. In a simulation surrogate, numerical library or engineering calculation, the key question is the same: which operation will consume the approximation, and what evidence shows that operation remains accurate enough?
CHAPTER 13 OF 14 · Transfer and continue learning
13. Changed-context transfer exercise
A constructed response on x in [2,6] is represented by p(t) = 5+T₁(t)−T₂(t)/2, where t = (x−4)/2. Predict the response at x = 5, its derivative there, and the integral across the full physical interval. State each transformation before calculating.
At x = 5, t = 1/2. Since T₂(1/2) = −1/2, the value is 5+1/2+1/4 = 23/4, or 5.75. Expanding the model gives p(t) = 11/2+t−t², so dp/dt = 1−2t. At t = 1/2 this derivative is zero; multiplying by dt/dx = 1/2 leaves the physical derivative zero.
The integral in t is 11−2/3 = 31/3. Because dx = 2dt, the physical integral is 62/3. These three answers use the same coefficient array but different operations. A correct value evaluator alone would not test the derivative or integration scaling.
Add uncertainty and change the domain
Suppose the three retained coefficients have independent known absolute error bounds of 0.01 each. Without assuming statistical cancellation, the function-value error caused by those coefficient errors is at most 0.03 throughout the standard interval. Independence is not needed for this worst-case sum; it would matter only if constructing a probabilistic error statement under additional assumptions.
Now ask for x = 8. This maps to t = 2, outside the model’s declared interval. The basis values are T₀ = 1, T₁ = 2 and T₂ = 7, so the same coefficient-error bounds give an allowance of 0.01(1+2+7) = 0.10 for that component of uncertainty. The in-interval 0.03 bound cannot simply be reused.
Finally, suppose the original function was known only through noisy measurements. The coefficient-error bounds would need justification from the fitting process; they cannot be invented from the number of displayed decimal places. Explain that change in evidence before recalculating. Transfer means recognizing which assumptions survive a new situation and which must be established again.
CHAPTER 14 OF 14 · Transfer and continue learning
14. Questions and a practical learning route
Is Chebyshev approximation the same as minimax approximation?
The terms are sometimes used loosely, so state the actual construction. A Chebyshev basis, interpolation at Chebyshev nodes, a truncated Chebyshev series and a minimax solution are distinct objects. Chapter seven gives an example where a truncated series is minimax and the corresponding three-node interpolant is not. The objective and evidence determine which claim is justified.
Can I use fixed experimental data?
Yes, a Chebyshev basis can be used in a fit to fixed data. However, you do not inherit every property of interpolation at deliberately chosen cosine-spaced nodes. Include measurement uncertainty, assess the sampling coverage and select a degree the information can support. Exact interpolation of every noisy measurement is not automatically the most useful model.
Does a tiny final coefficient prove the error is tiny?
No. A small computed tail is useful diagnostic evidence, but unresolved components may alias into lower terms or disappear at the nodes. Compare degrees, use independent queries and seek mathematical bounds when a guaranteed tolerance is required. Keep the strength of the conclusion proportional to the strength of those checks.
What should I practise first?
Derive T₂, T₃ and T₄ from the recurrence. Map three physical intervals to [−1,1] and check their endpoints. Recover the coefficients of the quadratic example from its three values. Evaluate the four-term series both directly and with Clenshaw. Then reproduce the t⁴ comparison and the aliasing examples before moving to an expensive target function.
For a small project, compare degrees three, seven and fifteen on one smooth function and one function with a corner. Use fresh test inputs and report observed errors, coefficient patterns and computational cost. Do not turn a sampled maximum into a formal bound. Explain where interval splitting helps and whether any downstream derivative requirement changes the conclusion.
A completed understanding includes the representation, sampling rule, evaluator, error measure and domain. Return to the Education parent with a concrete test of transfer: can you diagnose a changed unit, a hidden oscillation or a new derivative requirement without treating the coefficient array as an answer that explains itself?